Every signal has a source.

Balance, coverage and agreement

The balance is an average of directional readings across independent source families. Every family has an equal declared prior weight. Adding more technical indicators cannot multiply the technical family's vote.

Coverage measures the share of expected inputs currently available, averaged across all 15 families. Neutral is a measured reading. Missing, stale and not-yet-formed are different states and do not cast a neutral vote.

Agreement is the fraction of directional families on the majority side. Neither balance nor agreement is a probability of making money.

A complete trading call is withheld below 70% coverage, without verified price and options inputs, or with fewer than five directional families.

The time on the label matters

Each input records its observation date, collection time, source hash, frequency and release time when supplied. Daily means a daily publication, not that every underlying dataset is updated every day.

CFTC positions are generally measured Tuesday and released Friday. Monthly credit data and weekly liquidity observations are lagged. They are not substitutes for live flows.

Daily, weekly and monthly views apply distinct change windows. They are contextual scenarios. Numeric return projections require separately evaluated models; a longer chart does not become a forecast.

Official public observations

• CFTC Traders in Financial Futures: futures-only asset-manager and leveraged-fund categories. Net contracts / open interest; not ETF holdings, dark-pool trades or known dealer intent.
• Federal Reserve / FRED: DGS2, DGS10, DGS30, SOFR, EFFR, WALCL, WTREGEN, RRPONTSYD, WRESBAL, DTWEXBGS, BUSLOANS, TOTBKCR. Original government observations, not proprietary market-price series.
• ECB reference exchange rates: same-date USD cross rates and changes calculated by The Daily Consensus. Daily fixes are not executable quotes.
• The GDELT Project: a bounded English-language financial-news batch. Finance themes and headline terms filter relevance; normalized headlines collapse syndication. Tone describes language, with bins below −1, within ±1 and above +1. It is not investor sentiment or a directional trading vote. Headlines remain attributed to their publishers and may include releases or topic-classification errors.
• BLS releases, FOMC meetings, Treasury auctions and OCC / OIC expirations: source dates and supplied Eastern times. Date-only events retain that precision. Expiration calendars do not supply gamma, positioning or implied volatility.
• International policy calendars from the Bank of England, ECB and Bank of Japan. Decision days are separate from minutes and opinions. Date-only events retain their London, Frankfurt or Tokyo date; no exact time is inferred.
• Current price charts and technical summaries are hosted by TradingView. Provider timestamps and delays apply. The market overview uses SPY as its named benchmark. These current displays do not contribute to saved consensus scores and are hidden when reading a past edition.
• Official Federal Reserve, ECB, Bank of England and Bank of Japan release feeds. Headlines are linked to their publishers. This is a selected policy-news sample, not the entire news cycle.

Calculation assumptions

Government-series changes are scaled by declared constants and capped to −1 to +1. Falling rates and a weaker broad dollar are assigned supportive priors; higher reserves and bank credit are assigned supportive priors. Relationships can reverse by market regime. These rules are not learned accuracy.

The liquidity chart is Fed assets minus Treasury General Account minus overnight reverse repo, converted to common units and joined on the same observation date. It is a limited US proxy, not total world liquidity.

Fibonacci uses a fixed trailing window. Opening range requires every one-minute observation in the defined interval. Black–Scholes calculations require valid inputs and European-option assumptions; a signed open-interest gamma proxy is not observed dealer inventory.

An immutable track record

As-issued snapshots preserve source observations, cutoff, engine version and a SHA-256 digest. Collection alone does not justify a new issue date. New source observations, corrections or a declared model change do.

Outcome measurement begins at the next eligible session open after publication and ends after 1, 5 or 21 sessions. Missing prices do not skip sessions. A verified exchange calendar and a consistent adjustment basis are required.

Abstentions do not count as wins. Mean forecast returns are not a portfolio equity curve; overlapping horizons and correlated instruments must not be presented as independent trading profits.

Learning has to earn its place

Each instrument and horizon needs at least 120 resolved observations before a candidate can be evaluated. Inputs are normalized on training data only. Labels crossing the validation or test boundary are purged.

Candidate selection uses validation data, with the final chronological test held aside. A candidate must beat the historical base rate, a 50/50 baseline and the incumbent by at least 0.01 Brier score. Calibration bins and sample counts accompany every result.

No model has yet accumulated that prospective evidence. The original historical report's conviction percentages are not a verified win rate. No simulated results are included in the public record.